Method and arrangement for soft-decision sphere decoding

Abstract

A method is provided for soft-decision sphere decoding for softbit computation which can be applied to all sphere decoding algorithms, in particular sphere decoding algorithms in any mimo ofdm receiver implementations. Complexity reduction is achieved by using an approximate of linear euclidean distances during the sphere decoding search. The method can be used in conjunction with mimo ofdm communication systems like LTE, WiMax and WLAN.

Description

CROSS REFERENCE TO RELATED APPLICATIONS

This application claims priority of European application No. 10189471.5 filed on Oct. 29, 2010, the entire contents of which is hereby incorporated by reference herein.

FIELD OF THE INVENTION

The invention relates to a method and an arrangement for soft-decision sphere decoding.

BACKGROUND OF THE INVENTION

The system model of MIMO OFDM systems using N_T transmit and N_R receive antennas can be described in the frequency domain for every OFDM subcarrier individually by the received signal vector y=[y₁, . . . , y_NR]^T, the N_R×N_T channel matrix H, the transmitted symbol x=[x₁, . . . , x_NT]^T, and a disturbance vector n=[n₁, . . . , n_NR]^T which represents the thermal noise on the receive antennas. The following equation then describes the transmission model:
y=H·x+n  (1)

The elements of the transmitted symbol vector x are complex valued QAM symbols taken from a QAM modulation e.g. 4-QAM, 16-QAM, or 64-QAM. Depending on the modulation alphabet, every QAM symbol is associated to a number of transmitted bits N_Bit, with

$N_{\mathrm{Bit}}=\{\begin{array}{cc}2& \mathrm{for}4-\mathrm{QAM}\\ 4& \mathrm{for}16-\mathrm{QAM}\\ 6& \mathrm{for}64-\mathrm{QAM}\end{array}$

The elements of the channel matrix h_i,j are also complex valued. They are estimated by the receiver.

At a certain stage of the signal processing chain the receiver computes softbits for every transmitted bit associated to the transmitted symbol vector x. Several methods are known for this purpose, with different error probabilities and different computational complexities. One near-optimal approach in terms of error probability is soft-decision sphere decoding.

A soft-decision sphere decoder takes the received signal vector y and the channel matrix H as input and outputs a softbit (i.e. a likelihood value) for every bit associated to x. When denoting the bits associated to x_j (the QAM symbols of the j-th transmit antenna) by [b_j,1, . . . , b_j,n, . . . , b_j,Nbit(j)], a softbit p_j,n is defined by the following Euclidean distances:
d_0,j,n²=min_x0,j,n{∥y−H·x_0,j,n,∥²}
d_1,j,n²=min_x1,j,n{∥y−H·x_1,j,n∥²}  (2)
wherein d_0,j,n² and d_1,j,n² are the minimum Euclidean distances between the received signal vector y and all possible combinations of transmit symbols x, with the restriction that x_0,j,n represents all those combinations of x for which the n-th bit of the j-th transmit antenna is zero. On the other hand, x_1,j,n represents all those combinations of x for which the n-th bit of the j-th transmit antenna is one. The softbit for the n-th bit of the j-th transmit antenna is given by
ρ_j,n=d₀²−d₁²  (3).

A straight-forward algorithm would have to consider all combinations of x in the above equations in order to compute the softbits for one OFDM subcarrier. Since this approach is computationally very intensive and implies an exponential complexity, soft-decision sphere decoding algorithms have been proposed as a way to simplify the search. The simplification is achieved by QR decomposition of the channel matrix H followed by a tree search.

QR decomposition decomposes the channel matrix H into a orthogonal rotation matrix Q and an upper triangular matrix R, such that H=Q·R. Since rotation by Q does not influence the Euclidean distances in the above equations, one can simplify the Euclidean distances d_0,j,n² and d_1,j,n² by

$\begin{array}{cc}{d}_{0,j,n}^2={\min }_{x_{0,j,n}}\left\{{y^{\prime }-R·x_{0,j,n}}^2\right\}{d}_{1,j,n}^2=\underset{x_{1,j,n}}{\min }\left\{{y^{\prime }-R·x_{1,j,n}}^2\right\}\mathrm{with}{y}^{\prime }=Q^H·y.& \left(4\right)\end{array}$

A second step of the sphere decoding algorithm is the tree search.

The Euclidean distance from above, d²=∥y′−R·x∥², can be separated into partial Euclidean distances p₁², . . . , p_NT² as follows:

$\begin{array}{cc}{d}^2={\left(\begin{array}{c}{y}_1^{\prime }\\ \dots \\ y_{N_T}^{\prime }\end{array}\right)-\left(\begin{array}{ccc}{r}_{11}& \dots & r_{1N_T}\\ 0& \dots & \dots \\ 0& 0& r_{N_T{N}_T}\end{array}\right)\left(\begin{array}{c}{x}_1\\ \dots \\ x_{N_T}\end{array}\right)}^2=p_1^2+\dots +p_{N_T}^2,\mathrm{with}& \left(5\right)\\ p_{N_T}^2={y_{N_T}^{\prime }-r_{N_T{N}_T}·x_{N_T}}^2& \left(6\right)\\ p_1^2={y_1^{\prime }-r_{11}·x_1-\dots -r_{1N_T}·x_{N_T}}^2.& \left(7\right)\end{array}$

The partial Euclidean distances separate the original Euclidean distance into N_T portions. Due to the upper triangular structure of the R matrix, the partial Euclidean distances also separate the distance computation from the possibly transmitted QAM symbols x₁, . . . , x_NT such that p_NT² only depends on the QAM symbol x_NT and is not dependent on x₁, . . . , x_NT−1. Also, p_NT−1² only depends on x_NT and x_NT−1, and is not dependent on x₁, . . . , x_NT−2. This kind of dependency separation is utilized by the sphere decoding tree search in order to find the “closest” possible transmit symbol vector x_min.

The sphere decoding tree search assumes a maximum Euclidean distance d_max² which is definitely smaller than the Euclidean distance of the “closest” transmit symbol vector x_min. If now the search would start by choosing a candidate for x_NT, the partial Euclidean distance p_NT² is determined. In case of p_NT²>d_max², all the Euclidean distances d² for all possible combinations of x₁, . . . , x_NT−1 (assuming the chosen x_NT) will also exceed the maximum search radius d_max². Therefore, the search can skip computing the partial Euclidean distance p₁², . . . , p_NT−1², and can continue with another candidate for x_NT.

This search procedure can be illustrated as a tree search as depicted in FIG. 1. The search tree consists of N_T levels, that correspond to the QAM symbols of the different transmit antennas. In FIG. 1 N_T=3 is assumed. Each tree node is associated to one possible QAM symbol x₁, . . . , x_NT. Therefore, the leave nodes of the tree represent all possible combinations of x.

In the example above, with p_NT²>d_max², after choosing a candidate for x_NT the complete sub-tree below the chosen x_NT would be skipped during the sphere search.

For finding the “closest” transmit symbol vector x, the maximum Euclidean distance d_max² is initialized with ∞ (infinity). This means, that the partial Euclidean distances never exceed the limit, and that the sphere search reaches the bottom level after N_T depth-first steps. The resulting Euclidean distance d² then provides an update of the maximum search distance d_max². The sphere search would now continue and try to update d_max² if the bottom level of the tree is reached and if the resulting Euclidean distance would shrink d_max².

The result of this search process is d_max² being the Euclidean distance according to the “closest” possible symbol vector x_min. If x_min is restricted to certain bits being 0 or 1, the search tree can be adopted accordingly such that the search tree is built upon QAM symbols which meet the respective restrictions.

FIG. 2 illustrates an improvement of the sphere search by ordering the sibling nodes at a tree level k by increasing partial Euclidean distances p_k².

In a case where the maximum search distance d_max² is exceeded at a tree level k (solid tree node) and the partial Euclidean distances p_k² are not ordered, the search would continue with the next candidate node (the respective QAM symbol x_k) on the same level (arrow “A”). However, if the nodes in the tree are ordered by increasing p_k², the search can continue with the next node at level k−1 (arrow “B”). This is, permissible simply because due to the ordering of the sibling nodes the next candidate at the same level k would also exceed the maximum search distance d_max². In this case, the sub-tree which is skipped during the sphere search is much larger, and thus search complexity is much lower. It will be understood from the above that ordering of the sibling nodes by increasing partial Euclidean distances is essential for any efficient sphere decoding algorithm.

As mentioned above, Euclidean distances have to be computed during the sphere decoding algorithm which are given by the following equation:
d²=∥y′−R·x∥²  (8).

These distances are used as a search metric in order to find the closest possible symbol vector x_min and its associated Euclidean distance.

However, the computation of the Euclidean distances always requires multiplications for calculating the squared absolute value of a vector z=[z₁, . . . , z_NR] having complex elements z_r.
z=y′−R·x  (9)
d²=∥z₁∥²+ . . . +∥z_NR∥²  (10)

For practical implementations multiplications always involve significant computational complexity. Furthermore, multiplications increase the bit-width requirements of the multiplication result.

An object of the invention therefore is to provide a sphere decoding search algorithm with reduced computational complexity.

SUMMARY OF THE INVENTION

According to the invention there is provided a method for soft-decision sphere decoding.

The inventive method is adapted for use in a MIMO OFDM receiver with two receive antennas and comprises the steps of: receiving a channel matrix H and a received signal vector y; decomposing the channel matrix H into an orthogonal rotation matrix Q and an upper triangular matrix R, such that H=Q·R; performing a tree search based on Euclidean distances d² given by d²=∥z∥² to find a symbol vector x_min having a best likelihood to correspond to a transmitted symbol x, with z=y′−R·x and y′=Q^H·y. According to the invention, the tree search step comprises determining and using a linear approximation of the square-root of the Euclidean distances which is expressed as
{tilde over (d)}=(16·a₁+5·(a₂+a₃)+4·a₄)/16,
wherein a₁, a₂, a₃, a₄ are absolute values of the real and imaginary parts of z₁ and z₂, ordered in descending order, such that a₁≧{a₂, a₃}≧a₄, with z₁ and z₂ being the complex valued elements of the vector z.

The invention also provides an arrangement for soft-decision sphere decoding for use in an MIMO OFDM receiver. Advantageously, the arrangement according to the invention exhibits very low complexity; in particular it does not comprise any multipliers.

By using linear distances and in particular a linear approximation of the square-root Euclidean distances instead of squared Euclidean distances, the novel approach provides for significantly reduced computational complexity. The linear approximation of the square-root of Euclidean distances according to the invention is devised such that any multiplication operations can be dispensed with for computing d. Thus, the invention provides a way to significantly reduce computational complexity for practical implementations. A further advantage is the limited bit-width requirement on distance computation.

The invention can be used in conjunction with MIMO OFDM communication systems like LTE, WiMax, and WLAN.

BRIEF DESCRIPTION OF THE DRAWINGS

Additional features and advantages of the present invention will be apparent from the following detailed description of specific embodiments which is given by way of example only and in which reference will be made to the accompanying drawings, wherein:

FIG. 1 illustrates a tree search scheme;

FIG. 2 illustrates an optimization of sphere search in the tree search of FIG. 1; and

FIG. 3 shows a block diagram of an arrangement for computing the approximate square-root Euclidean distance according to the invention.

DETAILED DESCRIPTION

As stated before, the search metric for the sphere decoding search is based on the Euclidean distances d² given by d²=∥y′−R·x∥².

Instead, the sphere decoding search algorithm according to the invention uses the square-root of the Euclidean distances d given by
d=√{square root over (∥y′−R·x∥²)}  (11).

In this case, the search for the closest possible symbol vectors x_min will lead to the same result. However, the minimum search metric at the end of the search will be d instead of d².

For softbit computation for the n-th bit of the j-th transmit antenna still the given equation must be fulfilled:
ρ_j,n=d_0,j,n²−d_1,j,n²  (12).

When using square-root Euclidean distances d for the sphere decoding search, the multiplication would then be required for calculating p_j,n instead upon calculating the search metric. However, the inventors have realized that in this case the overall complexity is still much lower than if Euclidean distances d² would be used during the sphere decoding search.

For the case of a MIMO OFDM system with 2 receive and 2 transmit antennas (N_T=2, N_R=2) the square-root Euclidean distance is given by
d=√{square root over (∥z∥²)},  (13)
which corresponds to
d=√{square root over (real(z₁)²+imag(z₁)²+real(z₂)²+imag(z₂)²)}{square root over (real(z₁)²+imag(z₁)²+real(z₂)²+imag(z₂)²)}{square root over (real(z₁)²+imag(z₁)²+real(z₂)²+imag(z₂)²)}{square root over (real(z₁)²+imag(z₁)²+real(z₂)²+imag(z₂)²)}  (14).

It is known from literature, Paul S. Heckbert (editor), Graphics Gems IV′ (IBM Version): IBM Version No. 4, Elsevier LTD, Oxford; Jun. 17, 1994), chapter 11.2, that such distance metric can be approximated by the following linear equation
{tilde over (d)}=0.9262·a₁+0.3836·a₂+0.2943·a₃+0.2482·a₄  (15),
wherein a₁, a₂, a₃, a₄ are the absolute values of the real and imaginary parts of z₁ and z₂, ordered in descending order, such that a₁≧a₂≧a₃≧a₄. The coefficients for the approximation have been optimized to minimize the maximum relative error between d and d².

The method of soft-decision sphere decoding according to the invention uses a modification of the above linear approximation of expression (15). This modification has been devised by the inventor with regard to a very simple implementation thereof in hardware:
{tilde over (d)}=(16·a₁+5·(a₂+a₃)+4·a₄)/16  (16).

This linear metric can be implemented by simple shift operations and additions, rather than multiplications. Furthermore, for the disclosed metric (16), a₂ and a₃ do not have to be sorted necessarily, which eliminates one sorting operation. For calculating d with satisfying accuracy, a complete ordering such that a₂≧a₃ is not required. So, the sorting follows a₁≧{a₂, a₃}≧a₄ only.

FIG. 3 shows a block diagram of an exemplary embodiment of an arrangement for determining the approximate square-root Euclidean distance {tilde over (d)} according to the approximative expression (16) of the invention.

Since the approximation only involves multiplications by constants, no real multiplication is needed for calculating {tilde over (d)}.

In detail, the arrangement of FIG. 3 comprises an absolute-value generator 10 for determining the absolute value of the real part of z₁, an absolute-value generator 12 for determining the absolute value of the imaginary part of z₁ an absolute-value generator 14 for determining the absolute value of the real part of z₂, and an absolute-value generator 16 for determining the absolute value of the imaginary part of z₂.

The arrangement further comprises a comparator 20 connected to both of absolute-value generators 10 and 12 to determine a higher and a lower one of the two absolute values therefrom and to output them as a maximum and a minimum value, respectively. Similarly, a comparator 22 is connected to both of absolute-value generators 14 and 16 to determine and output a maximum and a minimum of the two absolute values therefrom.

A comparator 24 is connected to a first output of comparator 20 and to a first output of comparator 22 to receive the respective maximum absolute values therefrom. Comparator 24 compares the two maximum values and determines the higher one thereof as the highest of all four absolute values, i.e. a₁. A comparator 26 is connected to a second output of comparator 20 and to a second output of comparator 22 to receive the respective minimum absolute values therefrom. Comparator 26 compares the two minimum values and determines the lower one thereof as the lowest of all four absolute values, i.e. a₄.

As mentioned before, for the linear approximation according to the invention as set forth in expression (16), a sorting operation for a₂ and a₃ can be dispensed with. Rather, satisfying accuracy of soft-decision sphere decoding is obtained by sorting the four absolute values according to a₁≧{a₂, a₃}≧a₄ as performed by comparators 20, 22, 24, and 26. An adder 30 is connected to comparators 24 and 26 to receive therefrom the two intermediate absolute values to add them up to obtain a sum of a₂ and a₃.

The arrangement of FIG. 3 further comprises bit shifters 40, 42, 44, and 60. Left-shift operations by n bits are indicated by “<<n”, and right-shift operations are indicated by “>>n”. As can be seen in the figure, bit shifter 40 is connected to comparator 24 to receive a₁ to subject it to a left shift operation by 4 bits to effect a multiplication of a₁ by 16. Bit shifter 42 is connected to adder 30 to receive therefrom the sum of a₂ and a₃ to subject it to a left shift operation by 2 bits which effects a multiplication of the sum by 4. Bit shifter 44 is connected to comparator 26 to receive a₄ to subject it to a left shift operation by 2 bits to effect a multiplication of a₄ by 4.

An adder 50 is connected to adder 30 and to each of bit shifters 40, 42, and 44 to receive the outputs therefrom to add them all up, i.e. adder 50 sums 16·a₁ and 4·(a₂+a₃), and (a₂+a₃), and 4·a₄. Bit shifter 60 subjects the output of adder 50 to a right shift operation by 4 bits to implement a division of the sum from adder 50 by 16, and outputs the result as {tilde over (d)}, according to expression (16).

The disclosed method and arrangement for soft-decision sphere decoding using linear distances as described above provides a solution for further complexity reduction of all sphere decoding search algorithms. It can be shown by simulations that the introduced approximation to the square-root Euclidean distances is accurate enough for the overall soft-decision sphere decoding algorithm.

Claims

1. A method for soft-decision sphere decoding for use in a mimo ofdm receiver having two receive antennas, comprising the steps of:

receiving a channel matrix h and a received signal vector y;

decomposing the channel matrix h into an orthogonal rotation matrix q and an upper triangular matrix r, such that H=Q·R;

performing a tree search based on euclidean distances d² given by d²=∥z∥² to find a symbol vector x_min having a best likelihood to correspond to a transmitted symbol x, with z=y′−R·x and y′=Q^h·y;

wherein the tree search step comprises determining a linear approximation of the square-root of the euclidean distances which is expressed as

{tilde over (d)}=(16·a₁+5·(a₂+a₃)+4·a₄)/16,

2. The method of claim 1, wherein the step of determining a linear approximation of the square-root of the euclidean distances comprises the sub-steps of:

(a) determining absolute values of the real and imaginary parts of z₁ and of the real and imaginary parts of z₂;

(b) ordering said absolute values according to their magnitude to define a₁, a₂, a₃, a₄ such that a₁≧{a₂, a₃}≧a₄;

(c) adding up a₂ and a₃ to obtain a sum a₂+a₃; and

(d) performing a left shift operation by 2 bits on said sum;

(e) performing a left shift operation by 4 bits on a₁;

(f) performing a left shift operation by 2 bits on a₄;

(g) adding up results of steps (c) to (f); and

(h) performing a right shift operation by 4 bits on result of step (g).

3. An arrangement for soft-decision sphere decoding for use in a mimo ofdm receiver having two receive antennas, said arrangement being adapted for determining a linear approximation of the square-root of the euclidean distances which is expressed as a

{tilde over (d)}=(16·a₁+5·(a₂+a₃)+4·a₄)/16,

wherein a₁, a₂, a₃, a₄ are absolute values of real and imaginary parts of z₁ and z₂, ordered in descending order, such that a₁≧{a₂, a₃}≧a₄, with z₁ and z₂ being complex valued elements of a vector z, with z=y′−R·x and y′=Q^h·y, wherein h is a channel matrix, q is an orthogonal rotation matrix, r is an upper triangular matrix such that H=Q·R, x is a transmitted symbol, and y is a received signal vector, for performing a tree search to find a symbol vector x_min having a best likelihood to correspond to a transmitted symbol, wherein the arrangement comprises:

a first absolute-value generator for determining absolute value of the real part of z₁;

a second absolute-value generator for determining absolute value of the imaginary part of z₁;

a third absolute-value generator for determining absolute value of the real part of z₂;

a fourth absolute-value generator for determining absolute value of the imaginary part of z₂;

means for ordering said first, second, third, and fourth absolute values according to their magnitude for defining a₁, a₂, a₃, a₄ such that a₁≧{a₂, a₃}≧a₄;

a first adder connected to said ordering means to receive therefrom and add a₂ and a₃;

a first bit shifter connected to said ordering means to receive therefrom a₁ to subject a₁ to a left shift operation by 4 bits;

a second bit shifter connected to said first adder to receive therefrom sum of a₂ and a₃ to subject the sum to a left shift operation by 2 bits;

a third bit shifter connected to said ordering means to receive therefrom a₄ to subject a₄ to a left shift operation by 2 bits;

a second adder connected to said first adder and to each of said first, second, and third bit shifters to receive outputs therefrom to add up the outputs; and

a fourth bit shifter connected to said second adder to receive an output therefrom and subject the output to a right shift operation by 4 bits, and to output result as {tilde over (d)}.

4. The arrangement of claim 3, wherein the means for ordering the absolute values according to their magnitude for defining a₁, a₂, a₃, a₄ comprise:

a first comparator connected to the first and second absolute-value generators to determine and output a maximum and a minimum of said first and second absolute values;

a second comparator connected to the third and fourth absolute-value generators to determine and output a maximum and a minimum of said third and fourth absolute values;

a third comparator connected to a first output of the first comparator and to a first output of the second comparator to receive each of the maximum absolute values therefrom, to define the higher absolute value thereof as a₁ and to define the lower absolute value thereof as a₂ or a₃; and

a fourth comparator connected to a second output of the first comparator and to a second output of the second comparator to receive each of the minimum absolute values therefrom, to define the higher absolute value thereof as a₂ or a₃, and to define the lower absolute value thereof as a₄.